The $0 Stack

1. Four things, not a shopping list

A build standard does not care what you build it on. It cares that the substrate can do four things: hold records, run something on a schedule, send from an address the recipient recognises, and show a person what is in the book so they can fix a row. That is the whole requirement. Microsoft 365 and Google Workspace are popular answers because they bundle all four and most businesses already pay for them, not because the standard names them.

If you pay for neither, you assemble the four yourself. It takes an afternoon and it costs nothing, with one asterisk covered in section 3.

The fourth requirement is the one people skip and the one that decides whether this survives contact with a real office. Something has to be openable by whoever actually chases the money. If the only way to see the debtor book is to read a database, you have not built a back office, you have built a script with a database attached, and the first time a customer disputes a figure nobody can answer them.

2. The stack, and what each piece actually gives you

Every figure below was read off the vendor’s own pricing or limits page on 20 September 2026. Free tiers move, so check before you commit to one. The point of the table is not the numbers, it is that every number is knowable in advance.

The fourth requirement, the surface a person opens, is the one piece you build rather than sign up for. It is a table, an editable row, a queue and a log. Budget real time for it. It is the part that gets skipped when somebody is building this alone at nine at night, and it is the part they regret skipping first.

3. What happens when it works

Nothing dramatic, which is the point. These tiers are free because at your size you cost these companies almost nothing, and that is the arrangement rather than a loophole: they carry the small users and charge the ones who grow. A business running a few hundred jobs a month is not an edge case anybody is watching.

When you outgrow one piece you pay for that piece and nothing else moves. Resend goes to $20 a month and fifty thousand emails. Cloudflare and Neon both have a first paid tier in the same neighbourhood. You upgrade one line at a time, at the point where the thing you built is already earning, which is the right order to spend money in.

One detail worth knowing rather than worrying about. Vercel is the exception on this page: Hobby is for non-commercial personal use, and their definition of commercial explicitly includes “receiving payment to create, update, or host the site”. Build it yourself and none of their five examples describes what you are doing. Pay somebody to build it and you are in the one they wrote down. Either way the answer is $20 a month to stop thinking about it, and Cloudflare carries no equivalent restriction, which is why it leads the table above.

4. Free tiers sleep, and one cron fixes it

This is the part that actually breaks these builds, and it has nothing to do with running out of anything. Free tiers are designed to shut down resources nobody is using, because that is how the economics work. A back-office workflow is, by design, almost never using anything. A debtor chase wakes up once a day, works for thirty seconds and goes back to sleep. To a billing system optimising for abandoned side projects, that is indistinguishable from an abandoned side project.

The fix is one line of design and it costs nothing: give the job to a scheduler that lives outside all of this, and let it do double duty. The same thing that wakes your workflow every morning is also the thing touching the database, so the database never sits idle long enough to be considered abandoned. One cron, two problems, and the second one solved as a side effect of the first.

Upstash QStash is the obvious pick because it retries and keeps a delivery log, and cron-job.org works and emails you when a run fails. Either is free at this size. Build for it by exposing one authenticated endpoint that kicks the work off, so the schedule is never welded to whoever is hosting you and you can change your mind later without touching the workflow.

One sleep problem does not yield to this, and it is worth knowing which. A GitHub Action that has been disabled for inactivity cannot be pinged awake, because only commits count as activity and the scheduled runs themselves do not. That is a reason to keep the schedule somewhere other than the repository, and it is another reason the repository should be private, where the rule does not apply at all.

Then the last question is who watches the watchman, because a scheduler that stops is just a quieter version of the same problem. Have the workflow write down every time it runs, and have something look at that timestamp and complain when it goes stale. It is a row and a check. None of the vendors will build it for you, because from where they are sitting nothing is wrong.

This is not a pattern we are recommending from a distance. It is how we run our own: private repositories, Upstash driving every schedule from outside the thing it is scheduling, and nothing left anywhere that is allowed to fall asleep on its own.

5. The row that is never $0

Everything above is code and storage. Nobody is harmed if you get it wrong and nobody regulates it. A machine that picks up a phone and talks to an Australian consumer is a different kind of object. It has to disclose that it is not a person. It has to respect the Do Not Call Register and the calling-hours rules of the state it is ringing. It has to keep a record of what was said that is still there and still trustworthy when somebody disputes it months later.

You can build all of that. People do. It is simply the most expensive part of the job by a wide margin, it is the part where getting it wrong has consequences that are not technical, and it is the part where the cost never approaches zero no matter how the rest of the stack is assembled. That is why it is the part we sell, and why every standard names it explicitly rather than letting you discover it halfway through.

The published standards mark this with a with_us verdict rather than a failure, because a build that is correct everywhere else should not be told it failed. Run the checks against your $0 build and the verdicts sort themselves into two piles: what you can finish, and what you cannot. The second pile is the bill, and you can see it before you talk to anyone.

6. How you know it works

Not by looking at it. Every standard publishes its checks as data and there is a harness that runs them, so a build on free tiers gets asserted exactly the same way as a build on a subscription:

npx @aipathway/conformance show booked-after-hours-build-standard

The harness drives your build through the same ports whatever it is hosted on, which is the reason this page can be relaxed about the substrate. A build on Cloudflare and a build on Microsoft 365 pass identical checks or fail them for identical reasons. Nobody has to take your word for it, including you.

7. Whether you should do this at all

Worth saying plainly, because this page could easily talk somebody into a project they did not want. Assembling this is an afternoon if you can code and a long weekend of frustration if you cannot. The running cost is near zero and the attention cost is not: something you built is something you maintain, and the failure modes in section 4 are yours to catch.

If what you actually want is the phone answered properly after hours, buy that and skip all of this. It works on its own and it does not need a stack behind it. Build the $0 version when you want the workflow shaped to how your business actually runs, or when you want to own the thing outright, or because you enjoy it. Those are all good reasons. Saving $20 a month is not one of them.